All notable changes to this project will be documented in this file.
- Added in optional parameter for specifying a base_image.
- Created an example notebook that walks the user through a transfer learning example using a GPU.
- Added in a
clear_cache
function which deletes all files within the tmpfiles directory.
- Updated readme and implementation guide.
- Long-term change: .tmpfiles/ subdirectory to .AutoMLOps-cache/
- Long-term change: removed
use_kfp_spec
from parameter lists and switched to determining this at run-time.
- Removed redundant code for cloudbuild config generation.
- Updated constants.py file to remove constants no longer being used.
- Fixed custom imports mismatch between kfp spec and custom automlops spec.
- Verified custom_training_jobs_specs works as intended.
- Refactored backend modules to be structured based on frameworks (e.g. kfp, tfx, etc.) and deployments (e.g. cloudbuild, github actions, etc.)
- Added some unit tests for the utils.py module.
- Moved unit tests to /tests directory.
- New interface for defining AutoMLOps components and Pipelines. Removed the need to call Jupyter cell decorators and replaced them with Python function based decorators instead.
- New feature updates allow for running AutoMLOps outside of Jupyter Notebooks.
- Examples for running AutoMLOps outside of a notebook, as well as an example inferencing pipeline.
- Faster build jobs
- Updated readme and implementation guide.
- Better logging
- Better handling of requirements.txt (remove dups, infer from pipreqs, option to input explicit versions, sorted order)
- Better versioning of models
- Bug that can change current working directory on failure with run_local=True.
Official release on PyPI.
Staging for PyPI.
- Cleaning up wheel and egg files from repo.
- Remove dist/ and build/ directories.
Added feature to allow for accelerated/distributed model training.
- Support for custom training job specs (specifies which resources to use for pipeline jobs).
- Updated readme and implementation guide.
- New custom_training_job_specs parameter.
- Changed workflow for PipelineBuilder
- Bug related to grep substring match for create_resources script.
Reworked process to submit jobs to cloud runner service.
- Cloud Tasks Queue; API enabling, queue creation, and task generation.
- run() workflow reduced; submission to cloud runner service now takes place as part of the cloudbuild script.
- Creation of schedule job is now part of the cloudbuild script.
- Removed submit_to_runner_svc.sh and create_scheduler.sh.
- Added support for vpc_connectors.
- Bug related to waiting for cloudbuild job to complete before submitting to cloud runner service.
- Bug related to elevated IAM privileges in order to authenticate before submitting to cloud runner service.
First major release